12
Scene Recognition through Visual Attention
and Image Features: A Comparison between
SIFT and SURF Approaches
Fernando López-García1, Xosé Ramón Fdez-Vidal2,
Xosé Manuel Pardo2 and Raquel Dosil2
1Universidad
2Universidade
Politécnica de Valencia
de Santiago de Compostela
Spain
1. Introduction
In this work we study how we can use a novel model of spatial saliency (visual attention)
combined with image features to significantly accelerate a scene recognition application
and, at the same time, preserve recognition performance. To do so, we use a mobile robotlike application where scene recognition is carried out through the use of image features to
characterize the different scenarios, and the Nearest Neighbor rule to carry out the
classification. SIFT and SURF are two recent and competitive alternatives to image local
featuring that we compare through extensive experimental work. Results from the
experiments show that SIFT features perform significantly better than SURF features
achieving important reductions in the size of the database of prototypes without significant
losses in recognition performance, and thus, accelerating scene recognition. Also, from the
experiments it is concluded that SURF features are less distinctive when using very large
databases of interest points, as it occurs in the present case.
Visual attention is the process by which the Human Visual System (HVS) is able to select
from a given scene regions of interest that contain salient information, and thus, reduce the
amount of information to be processed (Treisman, 1980; Koch, 1985). In the last decade,
several computational models biologically motivated have been released to implement
visual attention in image and video processing (Itti, 2000; García-Díaz, 2008). Visual
attention has also been used to improve object recognition and scene analysis (Bonaiuto,
2005; Walther, 2005). In this chapter, we study the utility of using a novel model of spatial
saliency to improve a scene recognition application by reducing the amount of prototypes
needed to carry out the classification task. The application is based on mobile robot-like
video sequences taken in indoor facilities formed by several rooms and halls. The aim is to
recognize the different scenarios in order to provide the mobile robot system with general
location data.
The visual attention approach is a novel model of bottom-up saliency that uses local phase
information of the input data where the statistic information of second order is deleted to
achieve a Retinoptical map of saliency. The proposed approach joints computational
mechanisms of the two hypotheses largely accepted in early vision: first, the efficient coding
www.intechopen.com
186
Object Recognition
(Barlow, 1961; Attneave, 1954), which postulates that the mission of the first stages of the
visual processing chain is to reduce the redundancy or predictability in the incoming data;
and second, in the visual cortex relevant attributes of the image are early detected using
local phase or energy analysis, such as edges of objects. At those points where these features
are located there is an alignment of the local phase of the Fourier harmonics (Phase
Congruency). The model of local energy to detect features (Morrone & Burr, 1988; Morrone
& Owens, 1987; Kovesi, 1999) is based on this idea and demonstrated its suitability for
perceptual appearance and image segmentation. Nevertheless, it is not able to prioritize the
features with regards to the visual saliency. This fact is illustrated in Figure 1, where the
input image is formed by bars that increment its orientation in steps of 10º from left to right
and top to bottom, except for the central bar that breaks this periodicity creating a pop-out
effect for the HVS.
a) Original Image.
b) Salience from PC.
c) Salience from our model.
Fig. 1. Saliency maps of the original image; (a) from Phase Congruency (b) and the proposed
model (c).
In Fig. 1b we see the map of saliency achieved using Kovesi’s model (Kovesi, 1999) based on
Phase Congruency (PC). It provides approximately equal weight to all features clearing
away the pop-out effect. We think the reason for that is the high redundancy in images
which implies correlations and thus Gaussianism in chromatic and spatial components. It is
known that to handle information about phase structure is equivalent to use non-Gaussian
information in data distribution (Hyvärinen et al., 2009). Thus, to focus in the information
that does not depend on covariances (local phase) it is necessary to reduce redundancy, that
is, to decorrelate the data. One way is through data whitening. Redundancy in RGB color
components is deleted through PCA and spatial redundancy is avoided using an strategy of
filter-based whitening in frequency domain. In Fig. 1c it is shown that this hypothesis works
making possible to prioritize the salience of visual features from local phase.
Scene recognition is performed using SIFT (Lowe, 2004) and SURF (Bay, 2008) for image
featuring (two different approaches that we compare) and the Nearest Neighbor rule for
classification. SIFT features are distinctive image features that are invariant to image scale
and rotation, and partially invariant to change in illumination and 3D viewpoint. They are
fast to compute and robust to disruptions due to occlusion, clutter or noise. SIFT features
have proven to be useful in many object recognition applications and currently they are
considered the state-of-the art for general purpose and real-world object learning and
recognition, together with SURF features. SURF is a robust image descriptor used in
www.intechopen.com
Scene Recognition through Visual Attention
and Image Features: A Comparison between SIFT and SURF Approaches
187
computer vision tasks like object recognition or 3D reconstruction. The standard version of
SURF is several times faster than SIFT and it is claimed by its authors to be more robust
against different image transformations than SIFT. However, the results of our experimental
work showed that SIFT features perform significantly better than SURF features. In
combination with saliency maps, SIFT features lead to drastic reductions in the number of
interest points introduced in the database of prototypes (used in 1-NN classification), also
achieving very good performance in scene recognition. Thus, since the computing costs of
classification are significantly reduced the scene recognition is accelerated.
The chapter is developed as follows. Next Section presents the model of spatial saliency. An
overview of the image featuring methods is provided in Section 3. Section 4 deals with the
scene recognition application. Experimental work and results are presented in Section 5.
Finally, Section 6 is devoted to conclusions.
2. Model of spatial saliency
Figure 2 shows a general flow diagram of the saliency model. Following we describe each
stage of the model.
2.1 Early stage
The goal of this initial stage is to delete the statistical information of second order in color
components (RGB) and spatial components (between pixels of each color component),
through different whitening processes.
The aim of the initial step in this stage is to provide the model with a color space that contains
a mechanism, biologically inspired, called short-term adaptation (Simoncelli & Olshausen, 2001;
Barlow & Foldiak, 1989), which main goal is to achieve a final synchronization in the adaptive
process that promotes the most useful aspects for later processing. For that, the color RGB
image is decomposed into three channels maximally decorrelated using Principal Component
Analysis (PCA). Nevertheless, we are not interested in reducing the color space dimension,
thus, we use a transformed space of the original dimension (3 color components). The first
component corresponds to opponents channel B/W and the remaining two correspond to
opponents similar to R/G and Y/B. However, the space, unlike the opponents space CIE-Lab,
is adapted to the specific statistic of the incoming image.
In a second step, the goal is to eliminate the spatial redundancy among the pixels in each
color channel. In this case, we use a filter-based strategic in frequency domain called
Spectral Whitening (SW). It is consequence of the Wiener-Khinchin theorem: “for a
stochastic process, the average power spectrum is the Fourier Transform of the autocorrelation
function”. Thus, a whitened image should have a flat power spectrum. This can be easily
achieved using an adaptive filter in the frequency domain that normalizes the spectrum of
the transformed Fourier corresponding to the incoming image I(x,y) in the following way:
(
)
n ωx , ω y =
ℑ ⎣⎡ I ( x , y ) ⎦⎤
ℑ ⎡⎣ I ( x , y ) ⎤⎦
=
(
)
f ( ωx , ω y )
f ωx , ω y
where ℑ[⋅] is the transformed Fourier and ωs = ωx2 + ωy2
(1.1)
is the spatial frequency.
Physically, SW is a redistribution of the spectrum energy that will achieve an enhancement
www.intechopen.com
188
Fig. 2. General diagram showing how the data flows through the model.
www.intechopen.com
Object Recognition
Scene Recognition through Visual Attention
and Image Features: A Comparison between SIFT and SURF Approaches
189
of the less redundant patterns. This whitening method was previously used in the saliency
model proposed by Guo et al. (Guo et al., 2008) which is based on global phase analysis.
2.2 Analysis stage
In this stage, it is analyzed the maximum alignment of the local phase of each pixel in each
whitened color channel weighted by the strength of the visual features in the analyzed scale.
Classical methodology to estimate the amplitude and phase in a 1D signal is the Analytic
Signal. However, the 2D version was achieved partially using a quadrature phase bank filter
(Gabor like filters), until the introduction of the Monogenic Signal by Felsberg & Sommer
(Felsberg & Sommer, 2001). Our model uses this last methodology that achieves a new 2D
analytic signal from the Riesz’s transform, which is the 2D isotropic extension of Hilbert’s
transform. Its representation in Fourier’s domain is a set of two simple filters in phasequadrature that are not selective neither in scale nor orientation:
( H (ω ,ω ) , H (ω ,ω )) = ⎜⎜⎜
⎛
1
x
y
2
x
y
⎝
ωx
ωx2 + ωy2
i,
⎞
i⎟
ωx2 + ωy2 ⎟⎟
⎠
ωy
(1.2)
The Monogenic Signal is a vector function of three components formed by the original
signal and two components achieved by convolving it with the filters of Riez’s transform,
that is:
f M ( x , y ) = ⎡⎣ f ( x , y ), f ( x , y ) ∗ h1 ( x , y ) , f ( x , y ) ∗ h2 ( x , y ) ⎤⎦
(1.3)
where h1(x,y) and h2(x,y) are the representations in the spatial domain of H1(ωx,ωy) and
H2(ωx,ωy) respectively. Because filters H1 and H2 are oriented in frequency domain but are
not selective in scale, commonly a Gaussian like band-pass filter is used to build scaled
versions of Riez’s filters. In our case, we used the following log-Gauss filter:
(
)
Gs ωx , ωy = e
2
⎛
⎜ log ⎛⎜ ω ⎞⎟
⎝ ωo ⎠
− ⎜⎜
2
⎛ k ⎞⎞
⎜ ⎛
⎜ 2 ⎜ log ⎝⎜ ωo ⎠⎟ ⎟
⎠
⎝ ⎝
⎞
⎟
⎟
⎟
⎟
⎟
⎠
(1.4)
where ω=(ωx,ωy) is the spatial frequency, ωo =( ωox, ωoy) is the central frequency of the filter
and k is the parameter that governs the bandwidth of the filter. If gs(x,y) is the spatial
representation of previous filter, the monogenic space of scales is built as follows:
f M ,s ( x , y ) = ⎡⎣ f ( x , y ) ∗ gs ( x , y ) , f ( x , y ) ∗ gs ( x , y ) ∗ h1 ( x , y ) , f ( x , y ) ∗ gs ( x , y ) ∗ h2 ( x , y ) ⎤⎦ =
= ⎡⎣ f s ( x , y ), h1,s ( x , y ), h2 ,s ( x , y )⎤⎦
(1.5)
The chosen bank of filters is formed by three scales (s=3) which central wavelengths were
distributed in 1 octave from the minimum wavelength (assigned to λ1=8 pixels), that is λi={8,
16, 32} pixels. The k parameter was fixed to achieve a bandwidth of 2 octaves in each filter in
order to obtain a good spectral coverage in the bank of filters. A simple implementation of
the monogenic signal, in the frequency domain, can be found in (Kovesi, 2000).
www.intechopen.com
190
Object Recognition
Once it is achieved the monogenic decomposition, the importance of each visual feature is
measured by maximizing in each pixel of the image and for all the scales, the level of local
phase alignment of the Fourier Harmonics, weighted by the strength of the visual structure
in each scale (measured as local energy f M ,i ( x , y ) ). We call this measure Weighted
Maximum Phase Alignment (WMAP), and is the following:
WMPA ( x , y ) = w fdn ⋅ maxsi= 1
{f
( x, y )
}
⋅ cosθi =
⎧
⎛ E ( x , y ) ⎞ ⎫⎪
(1.6)
⎪
⎟⎬ =
= w fdn ⋅ maxsi= 1 ⎨ f M ,i ( x , y )i⎜ local
⎜ Elocal ( x , y ) ⎟ ⎪
⎪⎩
⎝
⎠⎭
⎧
⎛
⎞⎫
⎪
s
s
⎜
⎟⎪
⎛ s
⎞
⎪
⎜
⎟⎪
⎜ ∑ f i ( x , y ), ∑ h1,i ( x , y ), ∑ h2,i ( x , y ) ⎟
⎪
⎜
⎟⎪
i =1
i =1
⎝ i =1
⎠
=w fdn ⋅ maxsi= 1 ⎨( f i ( x , y ), h1,i ( x , y ), h2 ,i ( x , y )) ⋅ ⎜
⎟⎬
2
2
2 ⎞
s
s
⎪
⎜ ⎛⎜ ⎛ s
⎞ ⎛
⎞ ⎛
⎞ ⎟ ⎟⎪
⎪
⎜ ⎜ ⎜ ∑ f i ( x , y ) ⎟ + ⎜ ∑ h 1,i ( x , y ) ⎟ + ⎜ ∑ h( x , y ) ⎟ ⎟ ⎟ ⎪
⎪
⎜ ⎜ ⎝ i =1
⎠ ⎝ i =1
⎠ ⎝ i =1
⎠ ⎟ ⎟⎪
⎠ ⎠⎭
⎝⎝
⎩
M ,i
where f M ,i ( x , y ) is the monogenic signal for the i-th scale and θi is the angle between vectors
f M ,i ( x , y ) and Elocal . This angle measures the deviation of the local phase in the monogenic
signal at the i-th scale respect to the local energy vector in pixel (x,y).
We are only interested on those pixels where local phase is congruent for the most of the
used scales. Thus, our measure must incorporate a factor that penalizes too narrow
frequency distributions. Factor wfdn is achieved as it was proposed by Kovesi (Kovesi, 1999)
for his measure of local Phase Congruency (PC).
2.3 Output stage
The final stage of the model has the aim of achieving a Retinoptic measure of the salience of
each pixel in the image. For that, we integrate in each pixel the WMPA(x,y) measures of
each color channel:
Saliency ( x , y ) = ∑ WMAP( x , y )
3
c =1
(1.7)
Finally, a smoothing is introduced by a Gaussian filter and also a normalization in order to
make easy to interpret the saliency map as a probability function to receive attention.
2.4 Computational complexity
The computational efficiency of the model is low due to the load introduced by the PCA
analysis, which grows lineally with the number of pixels in the image (N) and cubically with
the number of components (color channels), O(M3+N M2). The number of components is low
and constant, M=3, thus, the asymptotic complexity depends on N. The computational
complexity of the model depends on the FFT (Fast Fourier Transform) complexity performed
in filtering processing. This complexity is O(N log(N)). On the other hand, the computational
timing of the model is low, by example, for an image of 512x384 pixels using an Intel Core2
Quad processor at 2.4 GHz and 4Gb of RAM memory, the algorithm takes 0.91 seconds. We
have to take into account that the algorithm is scientific software programmed in MATLAB.
www.intechopen.com
Scene Recognition through Visual Attention
and Image Features: A Comparison between SIFT and SURF Approaches
191
3. Image features
SIFT and SURF belong to a set of methods aimed to detect and describe local features in
images. Among these methods we can found (Mikolajczyk, 2005): shape context, steerable
filters, PCA-SIFT, differential invariants, spin images, complex filters, moment invariants
and gradient location and orientation histograms (GLOH). Nevertheless, SIFT and SURF
have captured recent attention of researchers working on applications like object
recognition, robot mapping and navigation, image stitching, 3D modeling, video tracking,
etc, being its comparison a current issue in literature (Bauer, 2007).
With regards to SIFT features, we used the Lowe´s algorithm (Lowe, 2004) which works as
follows. To identify the interest points (keypoints), scale space extrema are found in a
difference-of-Gaussian (DoG) function convolved with the image. The extremas are found
by comparing each point with its neighbors in the current image and adjacent scales. Points
are selected as candidate keypoint locations if they are the maximum or minimum value in
their neighborhood. Then image gradients and orientations, at each pixel of the Gaussian
convolved image at each scale, are computed. For each key location an orientation,
determined by the peak of a histogram of previously computed neighborhood orientations,
is assigned. Once the orientation, scale, and location of the keypoints have been computed,
invariance to these values is achieved by computing the keypoint local feature descriptors
relative to them. Local feature descriptors are 128-dimensional vectors obtained from the
pre-computed image orientations and gradients around the keypoints.
SURF features (Bay, 2008) are based on sums of 2D Haar wavelet responses and make a very
efficient use of integral images to speed-up the process. As basic image descriptors they use
a Haar wavelet approximation of the determinant of Hessian blob detector. There are two
versions: the standard version which uses a descriptor vector of 64 components (SURF-64),
and the extended version which uses 128 components (SURF-128). SURF are robust image
features partly inspired by SIFT, being the standard version of SURF several times faster
than SIFT. SURF features provide significantly less keypoints than SIFT, approximately the
half of them (see Figure 3).
Fig. 3. SIFT (left) and SURF (right) keypoints computed for the same frame.
www.intechopen.com
192
Object Recognition
4. Scene recognition application
Scene recognition is related with the recognition of general scenarios rather than local
objects. This approach is useful in many applications such as mobile robot navigation, image
retrieval, extraction of contextual information for object recognition, and even to provide
access to tourist information using camera phones. In our case, we are interested in
recognize a set of different scenarios which are part of university facilities formed by four
class rooms and three halls. The final aim is to provide general location data useful for the
navigation of a mobile robot system. Scene recognition is commonly performed using
generic image features that try to collect enough information to be able to distinguish among
the different scenarios. For this purpose we used SIFT and SURF alternatives.
To compute the SIFT features we used the original code by Lowe (http://people.cs.ubc.ca/
lowe/keypoints/). We also used the original code for SURF features by Bay et al
(http://www.vision.ee.ethz.ch/~surf/). To carry out the classification task we used the 1NN rule, which is a simple classification approach but fast to compute and robust. For the 1NN approach, we need to build previously a database of prototypes that will collect the
recognition knowledge of the classifier. These prototypes are a set of labelled SIFT/SURF
keypoints obtained from the training frames. The class of the keypoints computed for a
specific training frame will be that previously assigned to this frame in an off-line
supervised labeling process. The database is then incorporated into the 1-NN classifier,
which uses the Euclidean distance to select the closest prototype to the test SIFT/SURF
keypoint being classified. The class of every test keypoint will be assigned to the class of the
closest prototype in the database, and finally, the class of the entire test frame will be that of
the majority of its keypoints.
5. Experiments and results
The experimental work consisted in a set of experiments carried out using four video
sequences taken in a robot-navigation manner. These video sequences were grabbed in an
university area covering several rooms and halls. Sequences were taken at 5 fps collecting a
total number of 2,174 frames (7:15 minutes) for the first sequence, 1,986 frames for the
second (6:37 minutes), 1,816 frames for the third (6:03 minutes) and 1,753 frames for the
fourth (5:50 minutes). First and third sequences were taken in a specific order of halls and
rooms: hall-1, room-1, hall-1, room-2, hall-1, room-3, hall-1, hall-2, hall-3, room-4, hall-3,
hall-2, hall-1. The second and fourth sequences were grabbed following the opposite order
to collect all possible viewpoints of the robot navigation through the facilities. In all the
experiments, we used the first and second sequences for training and the third and fourth
for testing.
In the first experiment we computed the SIFT keypoints for all the frames of the training
video sequences. Then, we labelled these keypoints with the corresponding frame class:
room-1, room-2, room-3, room-4, hall-1, hall-2 or hall-3. The whole set of labelled keypoints
formed itself the database of prototypes to be used by the 1-NN classifier. For each frame of
the testing sequences their corresponding SIFT keypoints were computed and classified. The
final class for the frame was set to the majority class among its keypoints. Very good
performance was achieved, 95.25% of correct classification of frames. However, an
important drawback was the computational cost of classification, which was high despite
the fact that 1-NN is known as a low cost classifier. This was due to the very large size of the
www.intechopen.com
Scene Recognition through Visual Attention
and Image Features: A Comparison between SIFT and SURF Approaches
193
database of prototypes formed by 1,170,215 samples. In the next experiment, we followed
the previous steps but using SURF features instead of SIFT. In this case, recognition results
were very bad achieving only 28.24% of recognition performance with SURF-128 features,
and 25.05% using SURF-64. In both SURF cases the size of the database of prototypes was of
415, 845.
Although there are well known techniques for NN classifiers to optimize the database of
prototypes (e.g. feature selection, feature extraction, condensing, editing) and also for the
acceleration of the classification computation (e.g. kd-trees), at this point we are interested in
the utility of using the saliency maps derived from the visual attention approach. The idea is
to achieve significant reductions of the original database by selecting in each training frame
only those keypoints that are included within the saliency map computed for this frame.
Also, in the testing frames only those keypoints lying within the saliency maps will be
considered for classification. Once the database is reduced in this way, optimizing
techniques could be used to achieve even further improvements.
Fig. 4. Saliency regions at threshold 0.250 and corresponding SIFT keypoins.
In next experiments we carried out the idea showed in previous paragraph, although we
wanted to explore more in-depth the possibilities of saliency maps. As it was commented,
saliency measures are set in a range between 0 and 1, thus, we can choose different levels
of saliency by simply using thresholds. We will be the least restrictive if we choose a
saliency > 0.0, and more restrictive if we choose higher levels (e.g. 0.125, 0.250, etc). We
planned to use eigth different saliency levels: 0.125, 0.250, 0.375, 0.500, 0.625, 0.750 and
0.875. For each saliency level we carried out the scene recognition experiment (see Figure
4) achieving the percentage of recognition performance, and the size of the database of
prototypes. Results using SIFT and SURF features are shown in Tables 1, 2 and 3 and
Figures 5, 6 and 7.
www.intechopen.com
194
Object Recognition
Recognition %
Database Size
Database Size %
Original
95.25
1,170,215
100.0
Saliency > 0.125
95.25
779,995
66.65
Saliency > 0.250
94.72
462,486
39.52
Saliency > 0.375
93.45
273,908
23.41
Saliency > 0.500
92.21
157,388
13.45
Saliency > 0.650
89.30
86,161
7.36
Saliency > 0.750
83.31
42,418
3.62
Saliency > 0.875
56.03
15,894
1.36
Table 1. Results achieved using original frames and saliency maps with SIFT features.
Recognition %
Database Size
Database Size %
Original
28.24
415,845
100.0
Saliency > 0.125
33.51
273,775
65.84
Saliency > 0.250
86.56
157,394
37.85
Saliency > 0.375
32.01
88,059
21.18
Saliency > 0.500
66.55
47,767
11.49
Saliency > 0.650
67.06
24,338
5.85
Saliency > 0.750
35.27
11,040
2.65
Saliency > 0.875
18.33
3,971
0.95
Table 2. Results achieved using original frames and saliency maps with SURF-128 features.
Recognition %
Database Size
Database Size %
Original
25.05
415,845
100.0
Saliency > 0.125
27.74
273,775
65.84
Saliency > 0.250
51.50
157,394
37.85
Saliency > 0.375
25.64
88,059
21.18
Saliency > 0.500
28.97
47,767
11.49
Saliency > 0.650
67.33
24,338
5.85
Saliency > 0.750
34.89
11,040
2.65
Saliency > 0.875
19.22
3,971
0.95
Table 3. Results achieved using original frames and saliency maps with SURF-64 features.
www.intechopen.com
Scene Recognition through Visual Attention
and Image Features: A Comparison between SIFT and SURF Approaches
Fig. 5. Graphical results of recognition and database size using SIFT features.
Fig. 6. Graphical results of recognition and database size using SURF-128 features.
www.intechopen.com
195
196
Object Recognition
Fig. 7. Graphical results of recognition and database size using SURF-64 features.
Experimental results show that although SURF features collect significantly less interest
points than SIFT features (approximately the half of them) their performance is not adequate
for the scene recognition application. However, SURF features have proven to be adequate,
and faster than SIFT features, in other applications (Bay, 2008). Another interesting result is
that recognition performance of SURF features shows an irregular behavior with the
saliency thresholds, in both cases, SURF-64 and SURF-128. A maximum peak of 86.56% is
reached at saliency level 0.250 in SURF-128, while recognition results provided by SURF-64
features are worse. When using no saliency maps and even with some less restrictive
thresholds, recognition results of SURF features are very bad. This means that SURF features
loose distinctiveness as more interest points are used. This fact does not occur in SIFT
features, thus, SIFT features present more distinctiveness than SURF features in very large
databases of interest points. The best results are achieved using SIFT features, which
combined with saliency maps can reduce the amount of prototypes in the database up to
one order of magnitude, while the recognition performance is held, e.g. saliency level 0.500
in Table 1 and Figure 5. In this case, the performance drops to 92.21% (only 3.04 points from
95.25%) while the database size is drastically reduced from 1,170,215 to 157,388 prototypes.
6. Conclusions
In this work, scene recognition is carried out using a novel biologically inspired approach to
visual attention in combination with local image features. SIFT and SURF approaches to
image featuring are compared. Experimental results show that despite SURF features imply
the use of less interest points the best performance corresponds by far to SIFT features. The
SIFT method achieves a 95.25% of performance on scene recognition in the best case, while
the SURF method only reaches 86.56%. Another important result is achieved when we use
the saliency maps from the visual attention approach in combination with SIFT features. In
this case, the database of prototypes, used in the classification task of scene recognition, can
www.intechopen.com
Scene Recognition through Visual Attention
and Image Features: A Comparison between SIFT and SURF Approaches
197
be drastically reduced (up to one order of magnitude) with a slightly drop in recognition
performance. Thus, the scene recognition application can be significantly speeded-up. In
addition, the experiments show that SURF features are less distinctive than SIFT features
when we use very large databases of interest points.
7. Acknowledgements
Work supported by the Ministry of Education and Science of the Spanish Government
(AVISTA-TIN2006-08447) and the Government of Galicia (PGIDIT07PXIB206028PR).
8. References
Attneave, F. (1954). Informational aspects of visual perception. Psychological Review, Vol. 61,
No. 3, pp. 183-193, ISSN 0033-295X (print), ISSN 1939-1471 (online).
Bauer, J.; Sünderhauf, N. & Protzel, P. (2007). Comparing Several Implementations of Two
Recently Published Feature Detectors. Proceedings of The International Conference on
Intelligent and Autonomous Systems, (IAV), Toulouse, France.
Barlow, H.B. (1961). Possible principles underlying the transformation of sensory messages.
Sensory Communication, pp. 217-234.
Barlow, H.B. & Foldiak, P. (1989). Adaptation and decorrelation in the cortex. In The
Computing Neuron, pp. 54-72, Addison-Wesley, ISBN 0-201-18348-X.
Bay, H.; Ess, A.; Tuytelaars, T. & Gool, L. V. (2008). Speeded-Up Robust Features (SURF).
Computer Vision and Image Understanding, Vol. 110, No. 3, pp. 346-359, ISSN 10773142.
Bonaiuto, J. J. & Itti, L. (2005). Combining Attention and Recognition for Rapid Scene
Analysis, Proceedings of the 2005 IEEE Computer Society Conference on Computer Vision
and Pattern Recognition (CVPR05), pp. 90-90, ISBN 978-88-89884-09-6, San Diego
(USA), June 2005, IEEE Computer Society.
Felsberg, M. & Sommer, G. (2001). The Monogenic Signal. IEEE Transactions on Signal
Processing, Vol. 49, No. 12, pp. 3136-3144, ISSN 1053-587X.
García-Díaz, A.; Fdez-Vidal, X. R.; Dosil, R. and Pardo, X. M. (2008). Local Energy
Variability as a Generic Measure of Bottom-Up Salience, In: Pattern Recognition
Techniques, Technology and Applications, Peng-Yeng Yin (Ed.), pp. 1–24 (Chapter 1),
In-Teh, ISBN 978-953-7619-24-4,Vienna.
Guo, C.L.; Ma, Q. & Zhang, L.M. (2008). Spatio-Temporal Saliency Detection Using Phase
Spectrum of Quaternion Fourier Transform. In: Proceedings of IEEE Conference on
Computer Vision and Pattern Recognition 2008 (CVPR’2008), Nº 220, IEEE, pp. 1-8.
ISBN 9781424422425.
Hyvärinen, A.; Hurri, J. & Hoyer, P.O. (2009). Natural Image Statistics. A probabilistic
approach to early computational vision. Springer. ISBN 978-1-84882-491-1.
Itti, L. & Koch, C. (2000). A Saliency-based Search Mechanism for Overt and Covert Shifts of
Visual Attention. Vision Research, Vol. 40, pp. 1489–1506, ISSN 0042-6989.
Koch, C. & Ullman, S. (1985). Shifts in Selective Visual Attention: Towards the Underlying
Neural Circuitry. Human Neurobiology, Vol. 4, No. 4, pp. 219-227, ISSN 0721-9075.
Kovesi, P.D. (1999). Image features from phase congruency. Videre: Journal of Computer
Vision Research. MIT Press. Vol. 1, No. 3.
Available from: http://mitpress.mit.edu/e-journals/Videre/001/v13.html
www.intechopen.com
198
Object Recognition
Kovesi, P.D. (2000). MATLAB and Octave Functions for Computer Vision and Image
Processing. School of Computer Science & Software Engineering, The University of
Western Australia.
Available from: http://www.csse.uwa.edu.au/~pk/research/matlabfns/
Lowe, D. G. (2004). Distinctive Image Features from Scale-invariant Keypoints. International
Journal of Computer Vision, Vol. 60, No. 2, pp. 91–110, ISSN 0920-5691 (print), ISSN
1573-1405 (online).
Mikolajczyk, K. & Schmid, C. (2005). A Performance Evaluation of Local Descriptors. IEEE
Transactions on Pattern Analysis and Machine Intelligence, Vol. 27, No. 10, pp. 16151630, ISSN 0162-8828.
Morrone, M.C. & Burr, D.C. (1988). Feature detection in human vision: A phase-dependent
energy model. Proceedings of the Royal Society of London B: Biological Sciences, Vol.
235, No. 1280 , pp. 221–245, ISSN 1471-2954.
Morrone, M.C. & Owens, R. (1987). Feature detection from local energy. Pattern Recognition
Letters, Vol. 6, No. 5, pp. 303–313, ISSN 0167-8655.
Simoncelli, E.P. & Olshausen, B.A. (2001). Natural image statistics and neural
representation. Annual Review of Neuroscience, Vol. 24, No. 1, pp. 1193-1216, ISSN
0147-006X.
Treisman, A. & Gelade, G. (1980). A Feature Integration Theory of Attention. Cognitive
Psychologhy, Vol. 12, pp. 97–136, ISSN 0010-0285.
Walther, D.; Rutishauser, U.; Koch, C. & Perona, P. (2005). Selective Visual Attention
Enables Learning and Recognition of Multiple Objects in Cluttered Scenes.
Computer Vision and Image Understanding, Vol. 100, pp. 1-63, ISSN 1077-3142.
www.intechopen.com
Object Recognition
Edited by Dr. Tam Phuong Cao
ISBN 978-953-307-222-7
Hard cover, 350 pages
Publisher InTech
Published online 01, April, 2011
Published in print edition April, 2011
Vision-based object recognition tasks are very familiar in our everyday activities, such as driving our car in the
correct lane. We do these tasks effortlessly in real-time. In the last decades, with the advancement of
computer technology, researchers and application developers are trying to mimic the human’s capability of
visually recognising. Such capability will allow machine to free human from boring or dangerous jobs.
How to reference
In order to correctly reference this scholarly work, feel free to copy and paste the following:
Fernando López-García, Xosé Ramón Fdez-Vidal, Xosé Manuel Pardo and Raquel Dosil (2011). Scene
Recognition through Visual Attention and Image Features: A Comparison between SIFT and SURF
Approaches, Object Recognition, Dr. Tam Phuong Cao (Ed.), ISBN: 978-953-307-222-7, InTech, Available
from: http://www.intechopen.com/books/object-recognition/scene-recognition-through-visual-attention-andimage-features-a-comparison-between-sift-and-surf-app
InTech Europe
University Campus STeP Ri
Slavka Krautzeka 83/A
51000 Rijeka, Croatia
Phone: +385 (51) 770 447
Fax: +385 (51) 686 166
www.intechopen.com
InTech China
Unit 405, Office Block, Hotel Equatorial Shanghai
No.65, Yan An Road (West), Shanghai, 200040, China
Phone: +86-21-62489820
Fax: +86-21-62489821